On the twenty-eighth of the month, a project engineer sits down to assemble the pay application, and the most consequential number on the page is the percent complete by line item on the AIA G703 continuation sheet, because that number, summed and certified on the G702, is a request for money. Get it right and the draw is clean. Overstate it, even by a few points on a few lines, and you have signed an application certifying work in place that is not in place, which is a payment problem and, in front of a lender or a surety, can become a fraud-exposure problem. Computer vision now reads the jobsite, OpenSpace and Disperse style 360 capture walked through the building, and tells you what is installed, which is exactly the input you want for the G703. This lesson is about how to use that vision read responsibly: it is a fast, comprehensive measurement of installed progress, but it is not a verdict on the pay app, because the vision has a ceiling, it misreads work in progress, and the number it produces is dollars under the dollars gate. By the end you will be able to design the computer-vision progress verification step so the vision drafts the percent-complete and the human verifies it against the superintendent's walk before it ever backs a draw.

Why Vision Reads the Jobsite and the Pay App Reads the Money

The pay application is the contractual instrument by which a contractor requests payment for work completed, and on a standard AIA contract it is built on two forms: the G702 application and certificate for payment, and the G703 continuation sheet that breaks the contract sum into a schedule of values and records, line by line, the work completed this period and to date. The percent-complete on each G703 line is the engine of the whole document: it determines how much is earned, how much is billed, and how much the owner pays. That is why the number is not a status report but a financial assertion, and why an inflated percent-complete is not a clerical slip but an overstatement of money owed. The contractor certifies on the G702 that the work covered by the application has been completed in accordance with the contract documents, so the percent-complete numbers feeding it carry that certification's weight.

Historically the percent-complete was estimated: field staff walked the work, judged the lines, and entered the numbers, which was slow, partial, and subjective, because a busy engineer estimated some lines from memory, rounded others, and could not walk every area before the billing deadline. Computer vision changes the measurement. Tools in the OpenSpace and Disperse mold capture the jobsite as 360 imagery, a hard-hat or fixed capture walked through the floors, then align that imagery to the floor plan and the model and read what is installed: drywall hung, conduit run, ductwork in place, tile set. The vision reads every captured area against the schedule of values and produces a per-line percent-complete faster and more complete than a human estimate ever was, reading the physical state of the jobsite from imagery at a scale no human walk matches.

So the architecture of the step is clear: the vision reads the jobsite and proposes the percent-complete, and the pay app reads the money. The vision is the measurement, comprehensive and fast; the G703 is the financial document the measurement feeds. The error this lesson exists to prevent is letting the vision's read flow straight into the G703 as the billed percent-complete, because then a measurement, with its known limits and failure modes, becomes a certified financial assertion without a human between the two. The discipline is to keep the vision as the draft and the human as the one who verifies it and owns the certification, because the number is dollars, and dollars is one of the five gates.

The Dollars Gate: An Overstated Percent-Complete Is a Payment, Not a Note

The program's five verification gates, design intent, code, contract authority, dollars, and life-safety, sort AI outputs by what they touch and how hard the human must verify before relying on them. The percent-complete on a pay app sits squarely behind the dollars gate, because the number is money: it determines the draw, which the owner pays, which the lender funds, which the surety stands behind. An output behind the dollars gate is not verified to a mood of confidence but to the standard the consequence demands, which for a pay app is high, because the document is certified and the overstatement has teeth. This is the gate-not-mood discipline: you do not back the vision read because it looks right, you back it because you have verified it to the standard the dollars gate requires.

Consider what an overstated percent-complete actually is. If the vision reads a drywall line at 80% and the truth is 60%, and that 80% flows onto the G703 and is certified on the G702, the contractor has billed for 20% of that line's value that is not in place. Over-billing on a certified pay application is not a rounding note to fix next month: it can be a breach of the payment terms, it can trigger a clawback when discovered, and in front of a federally backed lender or a payment-and-performance surety it can become a false-claims or fraud question, because the certification asserted completion that did not exist. The asymmetry runs opposite to the direction a busy team might drift: under-billing costs the contractor cash flow, which is recoverable next period, while over-billing creates a certified false assertion, which is not cleanly recoverable and carries legal and relationship exposure. So the false-positive on the high side, the vision saying installed when it is not, is the dangerous error, and the verification must hunt it.

A percent-complete is not a status report, it is a request for money under a signed certification, so a computer-vision read that overstates installed work is not a note to correct next month, it is an over-billing the human must catch before the draw goes out.

This is why the cardinal rule, in its dollars form, governs this step: verify before pay-app. The vision read does not become the billed percent-complete by default and get corrected if someone notices; it is a draft a responsible human verifies against ground truth and then certifies, or corrects and certifies. Speed and comprehensiveness are exactly what make an unverified read dangerous: it can put a wrong number on every line of a hundred-line G703 in minutes, and the human who certifies the G702 owns every one of those numbers regardless of how they were produced. The dollars gate does not care that the vision drafted the number; it cares that the number is right, and the responsible-charge human is accountable for that.

The 90% Limit and the Slab-in-Progress Misread

Computer vision on the jobsite is powerful, and it is also bounded, and the boundary is where the failures live. The practical ceiling: vision reads installed, finished, visible work well and reads ambiguous, in-progress, or concealed work poorly. The vision sees what the camera can see and recognizes what its model was trained to recognize, so a clearly hung sheet of drywall, a run of conduit on an open ceiling, a set tile floor read reliably. But the vision does not see inside a wall once it is closed, does not know whether the slab it is looking at has its rebar tied and inspected or is bare formwork, and does not reliably distinguish work that is truly complete from work that merely looks complete from the camera's angle. The vision gets you most of the way, the easy comprehensive read; the last stretch, the ambiguous and concealed and in-progress, is where it is unreliable and where the human's judgment is irreplaceable.

The canonical failure mode is the slab-in-progress misread. The capture walks a deck where a slab pour is underway: forms are set, rebar is partially placed, some sections are poured and some are open. The vision, reading the imagery against the schedule of values, sees a slab area and reads it as substantially placed, returning a high percent-complete for the concrete line, because from the camera's view there is concrete and formwork that looks like a slab. But the slab is in progress: rebar is not fully tied, the pour is partial, the inspection that releases the next stage has not happened. The vision read the visible state and inferred completion, and the inference was wrong, because in-progress work is precisely the case the vision reads poorly. If that high percent-complete flows onto the G703, the contractor bills for concrete that is not in place, an over-billing on a structural line, and the dollars-gate failure is realized.

This failure is not a tooling defect to be patched away; it is a structural property of reading physical progress from imagery. Vision infers state from appearance, and appearance and completion diverge most exactly for in-progress and concealed work, which is the work that matters most on a pay app because it is where value is being put in place this period. The vision is least reliable precisely on the lines that are actively changing, which are the lines the current pay app is billing. That is why the human's verification cannot be a spot-check of a few finished lines; it has to concentrate on the in-progress and concealed work where the vision's read is weakest and the billing is most active, exactly where the superintendent's knowledge of the actual state of the work lives.

Verify the Vision Read Against the Superintendent's Walk

The ground truth for what is actually installed is not the imagery and not the vision's inference from it; it is the superintendent's knowledge of the work, confirmed by a walk. The superintendent walks the job daily, knows which slabs are poured and inspected and which are formed-and-waiting, knows what is behind the walls that closed last week, knows that the ductwork the vision read as installed is hung but not yet connected and tested. That knowledge is the ground truth, the actual state of the work rather than an inference from appearance. So the verification step is concrete: before the vision's percent-complete backs a pay app, the superintendent walks the work, or reviews the vision read against their walk, and confirms, corrects, or rejects the read line by line, concentrating on the in-progress and concealed lines where the vision is weakest.

This is decide-then-draft inverted into verify-then-certify, the cardinal rule's pay-app form made operational. The vision drafts the percent-complete comprehensively and fast, saving the superintendent from estimating every line from scratch. But the superintendent does not certify the vision's draft; they verify it against their walk and own the result. Where the vision and the walk agree, the read stands. Where they disagree, the slab the vision read at 70% that the super knows is 40% because the pour is partial, the super's walk governs, because the super's knowledge of the actual work is the ground truth and the vision's inference is the draft. The number on the G703 is the verified number, the super's, not the vision's, and the human who certifies the G702 is certifying their own verified read, which is what the dollars gate and the responsible-charge principle require.

The sequencing matters: the verification happens before the pay app, not after the draw. It is tempting, under a billing deadline, to bill the vision's read and reconcile later, but that inverts the cardinal rule into bill-then-verify, the posture that puts a certified overstatement out the door. The verification is a precondition of the certification, not a follow-up to it, because once the G702 is signed and the draw requested, the overstatement is a realized financial assertion, not a draft to correct. So the super's walk gates the pay app: no line's vision read backs the billing until the super has verified it, with the verification effort proportioned to where the vision is weakest and the dollars are largest.

Proportion the Verification to the Line's Consequence

Not every G703 line carries the same risk, and the verification should be proportioned to the consequence, which is the consequence-proportioned verification discipline applied to the schedule of values. A high-value structural line, the concrete, the steel, where the slab-in-progress misread lives and where a few points of overstatement is a large dollar figure, demands close verification: the super confirms the actual placement and inspection state against the vision read, line by line. A small finishes line, a low-value item the vision reads cleanly because it is visible and finished, can be verified more lightly, because the dollars at stake and the vision's reliability both favor it. The point is not to verify everything to the same depth, which would squander the vision's efficiency, but to concentrate the verification where the product of dollar value and vision unreliability is highest.

This proportioning is how the step keeps the vision's efficiency while controlling the dollars-gate risk. The vision's comprehensive read does the bulk measurement across all the lines, and the human's verification concentrates on the lines that matter most: high-value, in-progress, concealed, the lines where appearance and completion are most likely to diverge. A flat policy, walk every line to the same depth, would be either too slow to meet the billing deadline or too shallow to catch the structural overstatement, so the proportioning is what makes the step both fast and safe. The superintendent's time goes where the risk is, guided by the vision's read of which lines are active and high-value, and the reliably read, low-money lines get a lighter touch.

The proportioning also clarifies the asymmetry on the high side. The error the verification must hunt is the overstatement, the vision reading installed when it is not, because that is the over-billing the dollars gate exists to prevent; an understatement costs the contractor cash flow this period but creates no certified false assertion, so it is the recoverable error. The verification is tuned to catch the overstatement, especially on the high-value structural lines where the slab-in-progress misread produces exactly that error, and the superintendent's walk is the instrument that catches it, because the super knows the actual placement state the vision can only infer from appearance.

The Applied Problem: Design the CV-to-Pay-App Verification Step

Here is the exercise. Design the computer-vision progress verification step that feeds the pay-app backup on the G702/G703. Specify the vision read: the 360 capture (OpenSpace/Disperse style) aligned to the schedule of values, producing a per-line percent-complete draft across all captured areas. Specify the ground-truth verification: the superintendent's walk that confirms, corrects, or rejects the vision read line by line, governing where it disagrees with the vision. Specify the proportioning: which lines get close verification (high-value, structural, in-progress, concealed, where the slab-in-progress misread lives) and which get a lighter touch (low-value, visible, finished). Specify the sequencing: the verification gates the pay app, before the draw, not after.

Produce two things. First, the verification-step design: the vision read as the draft, the super's walk as the ground truth and the governing authority on disagreement, the consequence-proportioned verification concentrated where dollar value and vision unreliability are highest, and the sequencing that puts verification before certification. Build it so a project could run the vision capture, get the per-line draft, walk the work, and produce a verified G703 the engineer can certify on the G702 without backing an overstatement. Second, the dollars-gate analysis: why the percent-complete is dollars and therefore behind the dollars gate, why the overstatement (the high-side false positive) is the dangerous error the verification must hunt, why the slab-in-progress misread is the canonical failure, and why the verification must precede the pay app rather than follow the draw.

The deliverable is the verification-step design and the dollars-gate analysis; the lasting product is a CV progress verification step that uses the vision's comprehensive read to draft the percent-complete fast while the superintendent's walk verifies it against the actual state of the work before it backs a pay app. The professional who masters this gets the vision's speed and coverage, every captured line read against the schedule of values in minutes, without inheriting the vision's failure modes into a certified financial document, because the human verifies the draft against ground truth, concentrates the verification where the dollars and unreliability are highest, and certifies a number they own. That is verify-before-pay-app made operational: the vision measures, the human verifies, and the G702 carries a certification the contractor can stand behind in front of the owner, the lender, and the surety.

Key Takeaways

• The percent-complete on the AIA G703 continuation sheet, summed and certified on the G702, is the engine of the pay application: it determines the draw, so it is a financial assertion under a signed certification, not a status report, which is why an overstated percent-complete is an over-billing, not a note to fix next month.
• Computer vision in the OpenSpace and Disperse mold reads 360 jobsite capture against the schedule of values and drafts a per-line percent-complete faster and more comprehensively than a human estimate, which is the computer-vision engine doing what it does best: reading the physical state of the jobsite from imagery at scale.
• The percent-complete sits behind the dollars gate, so it is verified to the standard the consequence demands, not to a mood of confidence (gate-not-mood): an overstatement that flows onto a certified pay app can become a clawback, a breach, or a false-claims question in front of a lender or surety.
• The vision has a practical ceiling: it reads visible, finished work reliably and ambiguous, in-progress, or concealed work poorly, and the canonical failure is the slab-in-progress misread, where the vision sees concrete and formwork and infers a high percent-complete for a slab that is only partially poured and not yet inspected.
• The high-side false positive, the vision reading installed when it is not, is the dangerous error and the one the verification must hunt, because an overstatement creates a certified false assertion that is not cleanly recoverable, while an understatement only costs recoverable cash flow this period.
• The ground truth is the superintendent's knowledge of the actual work, confirmed by a walk: the vision drafts the percent-complete, the super verifies it line by line, and where they disagree the super's walk governs, so the number on the G703 is the verified number the certifying engineer owns, which is the responsible-charge principle and verify-before-pay-app made operational.
• The verification is consequence-proportioned: close verification on the high-value, structural, in-progress, and concealed lines where dollar value and vision unreliability are highest, a lighter touch on the low-value visible finished lines the vision reads cleanly, so the step keeps the vision's efficiency while controlling the dollars-gate risk.
• The sequencing is load-bearing: the verification gates the pay app before the draw, never after, because bill-then-reconcile inverts the cardinal rule and puts a certified overstatement out the door, whereas verify-then-certify keeps the financial assertion one the contractor can stand behind.